Robots at Brown University can copy Mona Lisa and imitate your handwriting.

It's hard to tell the truth from the truth.

Editor's Note: This article is from the Wechat Public Number "Big Data Digest" (ID: Big Data Digest), author's Abstract bacteria, 36 krypton authorized release.

Robots at Brown University can copy Mona Lisa and imitate your handwriting.

Big Data Abstracts

Author: Song Xinyi

Written language is an important characteristic that distinguishes human beings from other creatures. Some creatures can talk to each other like human beings, but only human beings can write characters with their own style: exquisite Chinese characters, complex flower letters, and each person's handwriting is unique.

But now, researchers at Brown University have developed new robots that can write and draw like humans by virtue of their own mechanics.

Robots at Brown University can copy Mona Lisa and imitate your handwriting.

Atunobu Kotani is an undergraduate at Brown University. With the help of his teacher, he created a deep learning algorithm that enables the robot to determine quite accurately which strokes to use and in which order to generate handwritten words and hand-drawn images.

The robot initially trained in Japanese, and then used the algorithm of "self-learning" to determine the order and position of strokes, copy words in languages it had never seen, write words including Chinese, English, Hindi, Korean, Greek and Urdu, and copy letters and strokes with 93% accuracy.

Robots can also copy simple line sketches, such as the Mona Lisa sketches. In the near future, robots can help us take notes and draw flowcharts in the workplace, and become our right assistant.

But the researchers also pointed out that there are still some shortcomings in robots, such as not being able to write from right to left like left-handed people.

Write and draw like a human being

Writing is a seemingly simple but actually elaborately designed process of complex action combination. When you write a word, you must know where to put your pen, hold it and draw a line, then pick up the pen and finish the line. Then when do you pick up your pen and draw another line?

To imitate other people's handwriting, we need to learn the factors of strokes, line width, writing habits, font inclination and so on in each character to copy its writing successfully.

Robots at Brown University can copy Mona Lisa and imitate your handwriting.

Just as robots have spent a long time learning to walk on two simple legs in human eyes, it is a very difficult goal for robots to learn to write in a way similar to people's writing. Machines cannot observe words in the same way as humans. For example, when people see the letter "A", they see three different lines, so it's easy to know to reproduce it with three strokes. But machines can only see a set of pixels, and Kotani's algorithm can help convert these pixels into strokes, so that machines can write letters as smoothly as humans.

Kotani's in-depth learning network for analyzing letters can reproduce a series of strokes needed to write letters, and then tell the robot to complete each stroke's action.

Because it is only copying, it does not need to understand the actual meaning behind the characters, so the robot can write any language, for example, it can write "hello" in ten different languages, which use a completely different set of characters behind them. The same method can also be applied to any type of line graph, such as simple sketches. This algorithm can help robots interact with people better in the future.

The machine learning system consists of two different models, one is the "global" model, which allows the robot to view the image as a whole, to help it determine the possible starting point of a particular word or role, and how to move to the next word. Another is the "local" model, which can help the robot to complete the current how to handle the pen in hand, that is, correct movement, placement and so on.

Self-taught: Drawing Mona Lisa from Brush to Brush

In order to train robots, the researchers provided a corpus of Japanese characters and provided information about how the combination strokes of characters should be written. Then it learns to create a model by itself, which can look at the pixels of characters and predict the position of each character stroke to begin with, and then give the position information that needs to be moved when drawing the stroke.

This enables robots to write languages they have never seen before, and the ability to teach themselves surprises researchers.

Robots at Brown University can copy Mona Lisa and imitate your handwriting.

Researchers wrote "Hello" on whiteboards in Hindi, Tamil and Italian, trying to confuse robots. As a result, the robot can observe each different language through machine vision, and then write a copy of each word separately, even if it could only write Japanese before. In addition to printing, the robot also wrote English cursive scripts, and the results were just as good.

Robots can also recognize immature words written by six-year-olds, and they can easily imitate the handwriting of a group of kindergarten children when they visit the laboratory.

Images can also be copied. After Kotani drew a rough sketch of Mona Lisa on the whiteboard, the robot also made a vivid copy. And unlike inkjet printers, this robot draws not line by line, but by imitating the brush strokes of human drawings.

Robots at Brown University can be copied

Please read the Chinese version for details.